Limit number of tokens per second for whisper. (#1958)

Otherwise, it spends lots of time in the loop if the EOT token
is not predicted.

Author: csukuangfj

sherpa-onnx/csrc/offline-recognizer-whisper-impl.h

@@ -131,7 +131,7 @@ class OfflineRecognizerWhisperImpl : public OfflineRecognizerImpl {
       auto cross_kv = model_->ForwardEncoder(std::move(mel));
 
       auto results = decoder_->Decode(std::move(cross_kv.first),
-                                      std::move(cross_kv.second));
+                                      std::move(cross_kv.second), num_frames);
 
       auto r = Convert(results[0], symbol_table_);
       s->SetResult(r);

sherpa-onnx/csrc/offline-whisper-decoder.h

@@ -33,7 +33,8 @@ class OfflineWhisperDecoder {
    * @return Return a vector of size `N` containing the decoded results.
    */
   virtual std::vector<OfflineWhisperDecoderResult> Decode(
-      Ort::Value n_layer_cross_k, Ort::Value n_layer_cross_v) = 0;
+      Ort::Value n_layer_cross_k, Ort::Value n_layer_cross_v,
+      int32_t num_feature_frames) = 0;
 
   virtual void SetConfig(const OfflineWhisperModelConfig &config) = 0;
 };

sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.cc

@@ -19,7 +19,8 @@ void OfflineWhisperGreedySearchDecoder::SetConfig(
 
 std::vector<OfflineWhisperDecoderResult>
 OfflineWhisperGreedySearchDecoder::Decode(Ort::Value cross_k,
-                                          Ort::Value cross_v) {
+                                          Ort::Value cross_v,
+                                          int32_t num_feature_frames) {
   auto memory_info =
       Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeDefault);
 
@@ -99,7 +100,12 @@ OfflineWhisperGreedySearchDecoder::Decode(Ort::Value cross_k,
   int32_t n_text_ctx = model_->TextCtx();
 
   std::vector<int32_t> predicted_tokens;
-  for (int32_t i = 0; i < n_text_ctx / 2; ++i) {
+
+  // assume at most 6 tokens per second
+  int32_t num_possible_tokens = num_feature_frames / 100 * 6;
+  num_possible_tokens = std::min<int32_t>(num_possible_tokens, n_text_ctx / 2);
+
+  for (int32_t i = 0; i < num_possible_tokens; ++i) {
     if (max_token_id == model_->EOT()) {
       break;
     }

sherpa-onnx/csrc/offline-whisper-greedy-search-decoder.h

@@ -18,8 +18,9 @@ class OfflineWhisperGreedySearchDecoder : public OfflineWhisperDecoder {
                                     OfflineWhisperModel *model)
       : config_(config), model_(model) {}
 
-  std::vector<OfflineWhisperDecoderResult> Decode(Ort::Value cross_k,
-                                                  Ort::Value cross_v) override;
+  std::vector<OfflineWhisperDecoderResult> Decode(
+      Ort::Value cross_k, Ort::Value cross_v,
+      int32_t num_feature_frames) override;
 
   void SetConfig(const OfflineWhisperModelConfig &config) override;